Independent knowledge work
GDPval-AA v2 Elo
Artificial Analysis same-harness result. Grok ranks third behind Opus 5 effort variants; its confidence interval overlaps Fable 5 and Qwen3.8 Max.
SPIRITT
Grok 4.6Grok 4.6 is built to keep going
SpaceXAI tuned Grok 4.6 for long-running agents, difficult coding and knowledge work, and stronger first passes on interactive and visual projects. Independent Artificial Analysis puts it at 61 on the Intelligence Index, with frontier-level agent performance at $2/M input and $6/M output. On SPIRITT, it gets the browser, terminal, files, tools, and durable context to turn that persistence into finished work.
Launch week on X
The official release, independent measurements, hands-on coding results, and an honest reminder that value does not mean best at everything.
What SpaceXAI changed
Grok 4.6 keeps the 500K context and headline price of Grok 4.5, then pushes harder on sustained execution: more agentic training, better self-checking on long trajectories, and more substantial visual and interactive first passes.
Built for the full trajectory
SpaceXAI trained Grok 4.6 across knowledge work, general coding, web development, CAD, kernel optimization, and other agentic environments. The launch report says longer runs increasingly showed the model testing and verifying its own work before moving on.
The API accepts text and images and returns text, supports low through xhigh reasoning effort, function calling and structured outputs, and has a 500,000-token context window. Standard pricing is $2/M input, $0.50/M cached input, and $6/M output; prompts at or above 200K move to $4/$1/$12 for the full request.
Long-running agentsCoding + knowledge workVisual first passes500K contextVision inputLow to xhigh reasoning
AA Intelligence Index v4.1.1
Artificial Analysis, August 12, 2026. Grok 4.6 high reaches the frontier at 61, tied with GPT-5.6 Sol max and behind Anthropic's two leaders. It is not the universal #1 model.
Where Grok 4.6 earns its place
The strongest case is the combination: near-top intelligence, sustained agent work, practical speed, and much lower operating cost. The coding boards also show where Sol and Fable still lead.
Independent tool use
Near frontierτ³-Banking
Artificial Analysis multi-turn customer-service tool-use evaluation. Grok is second in the displayed frontier pair.
Independent terminal work
Terminal-Bench v2.1
Artificial Analysis: 88.4%, in line with the leaders on v2.1. Do not compare this directly with the much harder v3.0 row below.
Independent long-horizon
Near frontierAA-Briefcase Elo
Grok sits at Fable-tier but behind Opus 5. It averaged about 53 turns and 0.5B input tokens versus Opus 5 max at about 103 turns and 2.0B.
Official coding
DeepSWE v1.1
SpaceXAI launch table. Grok improves sharply over 4.5, but Sol and Fable still lead this long-horizon software-engineering benchmark.
Official terminal caveat
Terminal-Bench v3.0
SpaceXAI launch table, harder v3.0 task set. Grok improves 10.3 points over 4.5 but remains clearly behind Sol and Fable.
Independent economics
WinnerCost per AA Intelligence task
Lower is better. Grok delivers 61 Index points at $0.84 per evaluated task, putting it on Artificial Analysis's price-performance Pareto frontier.
API speed
WinnerOutput tokens / sec
Artificial Analysis model-level snapshots, August 13, 2026. Figures represent first-party API performance and may change with the rolling measurement window.
Sources: Artificial Analysis Grok 4.6 analysis and live model/provider pages; SpaceXAI's August 12 launch report, API documentation, and model card; Arena.ai, Cognition, and verified direct X posts. Independent and vendor-reported results are labeled separately. SpaceXAI's competitor rows use best self-reported or public results rather than one controlled rerun. Grok 4.6 is compelling for supervised, cost-sensitive agent work; use your own evals for consequential or low-oversight workflows.
How It Works
From a fresh workspace to Grok 4.6 staying with the task across tools, files, and verification loops
01
Give it a real workspace
Open a SPIRITT workspace with the browser, terminal, files, integrations, and memory already available. Grok 4.6's long-trajectory training matters when it can inspect evidence and act, not only answer in a chat box.
02
Pick Grok 4.6
Choose Grok 4.6 in the model picker. Use high for the documented frontier operating point, then raise or lower reasoning effort when the consequence, latency, and cost of the task justify it.
03
Set a finish line and let it work
Give it the objective, source material, constraints, and acceptance checks. Grok can research, code, use tools, and iterate through verification steps; you keep approvals and final judgment where mistakes would be expensive.
Put Grok 4.6 on work that takes more than one step
A full cloud workspace, browser, terminal, tools, files, and memory. Pick Grok 4.6 and give it a finish line.
Questions
Frequently asked questions
01What is Grok 4.6?+
Grok 4.6 is SpaceXAI's August 2026 frontier model for long-running agents, coding, knowledge work, and interactive or visual projects. The API model ID is grok-4-6. It accepts text and images, outputs text, supports a 500,000-token context window, and offers reasoning effort from low through xhigh.
02How is Grok 4.6 different from Grok 4.5?+
It keeps the same headline $2/M input and $6/M output pricing while adding a longer supplemental training run, regenerated reasoning traces, more agentic reinforcement learning, and broader technical training. Artificial Analysis measured a five-point Intelligence Index gain, from 56 to 61, and SpaceXAI reports stronger self-testing and visual first passes.
03What is Grok 4.6 best for?+
Sustained research, repository work, root-cause analysis, tool-heavy knowledge workflows, and product prototypes where you want a substantial first version quickly. Its strongest buying case is frontier-adjacent capability at a much lower cost than premium peers, not a claim that it wins every benchmark.
04What are the main limitations of Grok 4.6?+
It trails leading models on official DeepSWE v1.1 and Terminal-Bench v3.0, its 500K context is half the million-token windows of some peers, and high-effort first-answer latency can be around 32 seconds in current provider measurements. Prompts at or above 200K tokens also double all token rates for that request. Keep human review for high-stakes outputs.
05How much does the Grok 4.6 API cost?+
For prompts below 200,000 tokens, SpaceXAI lists $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens. At or above 200,000 prompt tokens, the whole request is billed at $4 input, $1 cached input, and $12 output per million tokens. Priority processing costs more.
06How can I access Grok 4.6 on SPIRITT?+
You can use Grok 4.6 through SPIRITT Models, or connect your existing Grok subscription from the Subscriptions tab and keep using that subscription's available budget. Either option lets Grok 4.6 work inside the same SPIRITT workspace with your browser, terminal, files, tools, integrations, and durable context.
07What is a good example of using Grok 4.6 well?+
Ask it to investigate a real product or codebase problem end to end: inspect the source files and live behavior, identify the root cause, implement a bounded fix, run tests, compare screenshots, and return the evidence. Give it explicit acceptance criteria so its self-checking has a concrete finish line.
08Where is Grok 4.6 best to test for full agentic capabilities?+
SPIRITT Workspaces. Pick Grok 4.6 in the model picker and run real work in a fully equipped cloud environment: tools, browser, files, terminal, memory, and durable sessions. That is where the model can act as an agent, not just chat.
Buy from builders who use what they sellBuilt using